Forecasting the broadcast volume of cultural communication assisted by short video technology and analyzing the influencing factors

Data pre-processing and correlation analysis

There may be non-conforming data in the acquired data, so it has to be pre-processed by removing weights, missing values and data normalization. In this paper, equation 1 is used to normalize the data. $$\begin{array}{}\text{(1)}& X=\frac{X-X_{min}}{X_{max}-X_{min}}\end{array}$$ Then the two text features, title and synopsis, are subdivided using the word division function and regular matching, and the results are filtered using the deactivated word list, and the countervectorizer method is called to convert the text data into 128-dimensional word vectors and represent the video text vectors with the mean value as the features in the prediction algorithm of the losers. After pre-processing, the final 6563 crawled video data are selected for the study, and the playback volume after 24h is used as the prediction label to explore the relevant factors affecting the video playback volume.

Feature selection

In order to reduce the computational dimension and improve the model prediction performance, feature selection is used to screen the features, this paper adopts the Pearson correlation coefficient for feature selection, for the feature variables x and y in the data can be obtained through experimentation for a number of groups of data, denoted as $(x_i,y_i)(i=1,2,3,\dots ,n)$ , which is calculated by the formula: $$\begin{array}{}\text{(2)}& P_{x,y}=\frac{cov(x,y)}{{\rho }_x{\rho }_y}=\frac{E((x-{\mu }_x)(y-{\mu }_y))}{{\rho }_x{\rho }_y}\end{array}$$

where ${\rho }_x=\sqrt{\frac{1}{n}\sum _{i=1}^n(x_i-{\mu }_x)}$, ${\mu }_x$ and ${\mu }_y$ denote the mean and variance of x, respectively, and the pear son correlation coefficient ranges from [-1, 1], and the correlation coefficient tends to -1 or 1 when the linear relationship of the features is enhanced. Pear son correlation coefficients of the feature variables and the predictor labels are shown in Figure 3.

Model selection

is an unsupervised machine learning algorithm, which is essentially a classification and prediction algorithm that builds Bagging integration and combines stochastic subspace based on decision trees [23]. The basic idea is to use boos trap sampling method to extract $n(n<N)$ samples from original sample data, repeat times and keep the sample capacity and the original dataset the same, and then in each decision tree node splitting, m sub-features are randomly extracted from all the features, and then select the optimal features from the sub-features as the split features; decision tree models $(f_1(X),f_2(X),\dots ,f_k(X))$ and predictions are obtained through training sessions, and the final prediction is decided by voting based on the results, and the model is represented as follows:

$$\begin{array}{}\text{(3)}& F(X)=\arg \underset{Y}{max}\sum _{i=1}^{k}I(f(X)=Y)\end{array}$$

Where $F(X)$ denotes the combined predictive model, $Y$ denotes the labelled features, $I(.)$ is the schematic function and X denotes the feature variables.

The XG Boost algorithm is based on the gradient boosting decision tree to optimize the loss function and feature selection, which can effectively construct the enhancement tree and run in parallel [24]. The idea of the algorithm is to use the latter model to correct the error generated by the former model, and continuously repeat to achieve the purpose of optimizing the objective function, the overall model and the objective function can be expressed as: $$\begin{array}{}\text{(4)}& {\hat{y}}_i=\sum _{k=1}^k{f}_k(X_i)\end{array}$$ $$\begin{array}{}\text{(5)}& Goal(\theta )=L(\theta )+\mathrm{\Omega }(\theta )\end{array}$$ Where, $i=1,2,L,n,n$ is the number of samples, $y_i$ is the predicted output, $K$ is the number of trees, $f_k(X_i)$ is the number of kth trees, $F$ is the set of all regression trees, and $L(\theta )=\sum _{i=1}^n(y_i-{\hat{y}}_i),\mathrm{\Omega }(\theta )=\sum _{k=1}^k\mathrm{\Omega }(f_k)$, denote the error term and the regularization term, respectively.

The LSTM neural network model is a special deformation of the recurrent neural network that improves on its tendency to have vanishing gradients and gradient explosions, allowing the network to achieve better results over longer data sequences [25].

Evaluation indicators

In order to further judge the strength of the model, this paper uses the goodness-of-fit $R^2$ to measure the model fitting accuracy and the mean absolute error (MAE) to measure the model prediction accuracy.

3. Experimental results and analyses

Prediction experiments for all candidate features

In this study, we use five-fold cross-validation to divide the training set and validation set, and the RF model uses Grid Search CV to achieve automatic parameter tuning to determine the optimal parameters of the model are $ma{x}_{depth}=3$, the number of iterations $n\mathrm{\_}es-timators=80$, respectively; the auxiliary parameters of the XG Boost model, the learning rate, the proportion of random samples, and the maximum height of the tree have a greater impact on the algorithm’s performance when the model is predicted. The performance of the XG Boost model has a large impact on the performance of the algorithm, and the number of iterations $n\mathrm{\_}estima-tors=30$, the task function is gamma, the learning rate $learning\mathrm{\_}rate=0.3$, the initial prediction score $base\mathrm{\_}score=0.5$, and the maximum height of the tree is 3. Comparisons of predicted and actual values between the RF model and the XG Boost model are shown in Figure [f4] and [f5], respectively.

The LSTM neural network model requires more parameters to be adjusted, the target variable is normalized in the input layer according to equation 1, the dropout is set to 0.01 in order to prevent overfitting, MAE is used as the loss function, Adam is used as the optimizer for the model, the data training is set to be performed for 5000 epochs, and each time, 16 training data are used to perform the forward and back propagation, and the model achieves very good results after 80 epochs. The model achieves very good results after 80 epochs and Figure 4 represents the prediction results of the LSTM neural network model on the validation set.

From Figure [f4], [f5] and 4, it can be seen that the curves of the predicted and true values of the RF model and the XG-Boost model are the closest to each other, and the error between the prediction result of the model and the true value is smaller, and the model fitting effect is better; the error between the prediction value and the true value of the LSTM neural network model is bigger, and the model fitting effect is poorer. As shown in Table 1, the RF model has a $R^2$ of 0.968 and a MAE of 215.917, while the LSTM model has a $R^2$ of 0.702, and the evaluation metrics show that the model has a higher prediction accuracy on the test set. From the evaluation indexes, it can be seen that the RF model has a better prediction effect than the XG Boost model and the LSTM neural network model in predicting the amount of playback of pop-up videos, while the LSTM neural network model has the worst performance.

Comparison experiment of different features under the optimal model

On the basis of the RF algorithm for predicting the playback volume of pop-up videos, we select the video features and creator features to predict the video playback volume, and explore whether the video features have the greatest impact on the playback volume by comparing the evaluation indexes; secondly, the feature variables can be divided into numerical and textual features, and we compare the results of the experiments of predicting the playback volume by using the numerical and textual features, and predicting the video playback volume by using all the features. Secondly, the feature variables can also be classified into numerical and textual features, comparing the experimental results of prediction using both numerical and textual features with those using all features, and analyzing whether the textual features of the video can improve the prediction effect and the degree of influence on the video playback. The experimental results of different features are shown in Figure [f7].

As can be seen from the figure, $R^2$ obtained from the prediction of the video’s own features is 0.966, and the MAE is 215.920, which is similar to the prediction of the candidate features, and the $R^2$ obtained from the prediction of the creator’s features is 0.223, which shows that its effect on the video playback is almost negligible, and it further indicates that the video’s own features have the greatest effect on the video playback. Secondly, comparing the prediction results of the evaluation indexes and candidate features predicted by numerical features and textual features respectively, it can be seen that the results predicted by the added textual features are obviously better than those predicted by the numerical features only, which indicates that the textual features affect the video playback volume to a certain extent and the influence is relatively small. Therefore, in the daily promotion process, pop-up video websites can focus on the video features and numerical features, which have a greater impact on the amount of playback, so as to promote the growth of video playback and improve economic benefits.

4. Discussion

Short videos are characterized by instant dissemination, short production cycle, wide range of contents, high netizen participation, free and flexible forms, easy sharing, etc. Carrying out service marketing activities relying on short videos can help to improve the knowledge, accessibility and acceptability of public cultural services, and is thus of great significance to enhancing the accessibility of public cultural services. Relying on the short video marketing “TRUST model” and combining the key elements of short video marketing of public cultural services, this study opens up a new marketing ecological perspective. Specifically, it can be laid out in the following five aspects to realize the whole process of user focusing, linking, attracting, expanding and retaining, and thus enhance the accessibility of public cultural services. and accessibility of public cultural services.

The object of short video marketing of public cultural services is users, and different types of users have different degrees of matching public digital cultural preferences and needs, which, if generalized, will result in the loss of some users or difficulty in attracting users. At the same time, short videos of public cultural services have wide and broad content, but a large number of different types of short videos of public cultural services are piled up, lacking in attributes such as recognition, characterization and specialization, thus making it difficult to attract the target group and maintain user stickiness. Therefore, only by deeply ploughing into the vertical segmentation of short videos of public cultural services and accurately focusing on the target audience can we improve the suitability of video content and users and achieve the goal of focusing on users. The specific measures are mainly divided into the following two aspects: First, customize the exclusive IP for short videos of public cultural services. Creating a number of exclusive IP for short videos on public cultural services, secure both the vertical consistency of a user’s IP and the short video content it publishes, as well as the IP’s focus on a particular niche. For example, “Fan Deng Book”, “Little Zhu History Lesson”, “Culture Reddit”, according to the IP, you can know its specific content, so as to focus on the target users. Second, customization the exclusive content of public cultural service short videos. User demand for public cultural service short videos is broadly divided into information demand, entertainment demand and emotional demand, respectively, for different needs to customization different attributes of public cultural service short video content and continuous output, accurate radiation of different needs of users [26, 27, 28].

The content of the marketing of public cultural services is the prerequisite and foundation for the development of public cultural services, as well as the key to building strong relationships with users. Therefore, the short video marketing of public cultural services needs to create high-quality content, optimize the content production from the perspective of showing the quality of the short video content of public cultural services and enhancing the users’ experience of the short video content of public cultural services, so that the users can realize the transformation from “recognition” to “approval”, thus linking the users. Firstly, show the sense of quality of public cultural service short video content. In the creation of short video content for public cultural services, attention should be paid to returning to the value of the content itself and driving quality with value. From the treasure trove of Chinese cultural resources to refine the subject matter, get inspiration, draw nutrients, the Chinese outstanding traditional culture of useful ideas, artistic value and the characteristics and requirements of the times combined, from a variety of perspectives on the depth of the interpretation of the public cultural content, thick plant the public cultural works of the “cultural content”, and continue to launch a deep, nurturing heart We will also continue to launch a series of short cultural videos with deep content that will cultivate people’s hearts and minds.

5. Conclusion

The B station pop-up video is selected as the research object, comparing the advantages and disadvantages of traditional machine learning methods and deep learning methods in the prediction of pop-up video playback, and different features are selected for comparative experiments to explore the key factors affecting the video playback. The experimental results show that the RF algorithm is better than the XG Boost algorithm and the LSTM neural network model in predicting on the data base of this study, and the error of the prediction made by the LSTM neural network is relatively large. Using the RF algorithm to select different features for prediction, comparing the experimental results, we found that the features of the pop-up video itself have the greatest impact on the video playback volume, and the video text features only affect the video playback volume to a certain extent and the impact is small. In the subsequent work, we will deeply explore the comment and pop-up contents of videos and perform sentiment analysis on them as feature variables to extend the prediction model, and track and monitor the video playback data within a certain period of time, using time series data and applying the corresponding model to make predictions, so as to deeply explore the influence mechanism of video playback and the prediction algorithm.

Funding information

This research was supported by the Higher Education of Social Science Research Program, Jiangsu Province, under the project code 2021SJA0552.